Detecting Moving Objects With Machine Learning

TL;DR

The chapter surveys machine learning approaches to detecting moving objects in astronomical imagery, focusing on streak detection, moving-point sources in sequences, and shift-n-stack methods, with convolutional neural networks as the dominant tool and attention to overfitting. It juxtaposes classic non-ML pipelines (image differencing, tracklet linking, and digital tracking) with ML-enabled techniques across surveys and objects, including KBMOD, ATLAS, ZTF, and MOA applications. The discussion covers clustering-based filtering, direct-image CNN classifiers, and regression networks for photometry, emphasizing training data quality, augmentation, and cross-survey generalization, while highlighting practical metrics and potential failure modes. The analysis extends to big-data contexts such as Euclid and LSST, addressing non-linear shift-n-stack scalability and proposing architectural directions to maximize discovery depth, efficiency, and reliability in future Solar System surveys, with a call for robust validation practices. $O(t^4)$ is cited as the scaling for non-linear shift-n-stack with time baseline, underscoring the need for advanced ML solutions to harness the full potential of upcoming datasets.

Abstract

The scientific study of the Solar System's minor bodies ultimately starts with a search for those bodies. This chapter presents a review of the use of machine learning techniques to find moving objects, both natural and artificial, in astronomical imagery. After a short review of the classical non-machine learning techniques that are historically used, I review the relatively nascent machine learning literature, which can broadly be summarized into three categories: streak detection, detection of moving point sources in image sequences, and detection of moving sources in shift and stack searches. In most cases, convolutional neural networks are utilized, which is the obvious choice given the imagery nature of the inputs. In this chapter I present two example networks: a Residual Network I designed which is in use in various shift and stack searches, and a convolutional neural network that was designed for prediction of source brightnesses and their uncertainties in those same shift-stacks. In discussion of the literature and example networks, I discuss various pitfalls with the use of machine learning techniques, including a discussion on the important issue of overfitting. I discuss various pitfall associated with the use of machine learning techniques, and what I consider best practices to follow in the application of machine learning to a new problem, including methods for the creation of robust training sets, validation, and training to avoid overfitting.

Figures (12)

• Figure 1: Top: Three images centred on a real Kuiper Belt Object discovered as part of the New Horizons Search for distant Kuiper Belt Objects Fraserlpsc2023. The source is just below the noise floor of the images. Bottom: The shift'n'stack sequence of $\sim100$ images, shifted at four different linear rates of motion of the same source shown above. The source's true rate of motion is 2.4 "/hr, and when the shift rate is near that value, the source becomes an easy detection.
• Figure 2: The custom architecture utilized by ChybaRabeendran2021 where each image of an ATLAS quad is passed through a pretrained ResNet-18 (labelled as "CNN"), and the confidence outputs are concatenated into a single 32-element vector that is passed through the classifier percepteron. Reproduced from Figure 6 of that work.
• Figure 3: Examples of the 8 different classes utilized in the classification search of ChybaRabeendran2021. The three real minor body classes are shown at right, with the 5 bogus classes resulting from detector and pipeline artifacts. Reproduced from Figure 3 of that work.
• Figure 4: The bi-FPN based network used by Duev2021 to identify comets in ZTF imagery. The network consists of two parts: an EfficientNet-B0 and biFPN section that creates multi-scale feature maps from the input imagery, and a CNN+percepteron head network that performs the binary classification and regression on comet position. Reproduced from Figure 3 of that work.
• Figure 5: The custom ResNet-style block invented and utilized by Cowan2023. Note the difference with the standard ResNet block which sums only one convolutional chain with the skip connection. Reproduced from Figure 9 of that work.